International Science Council

• Agreement that AI governance must address not only risks to individuals and societies, but structural risks to the systems through which humanity generates and validates knowledge • Recognition that science requires dedicated governance attention, distinct from general AI ethics or safety frameworks • Commitment to inclusive participation that gives meaningful voice to scientific communities, research institutions, and science funders, including from the Global South • Establishment of mechanisms for ongoing structured dialogue between the AI governance community and the international science system, including national academies, science unions, and intergovernmental science bodies • A shared understanding that governance decisions made now will shape the epistemic infrastructure of the next generation and that this deserves long-term, futures-oriented deliberation, not only near-term risk mitigation

• Safe, secure and trustworthy AI
• AI capacity-building
• Transparency, accountability, and human oversight
• Open-source software, open data and open AI models

• Safe, secure and trustworthy AI: Trust in AI-assisted science is inseparable from trust in science itself. Governance frameworks must ensure that AI use in research meets the standards of transparency and reliability that underpin scientific credibility and public trust in evidence • AI capacity-building: Capacity gaps in AI are not only technical but also encompass the ability of scientific institutions, especially in the Global South, to participate in and shape the AI systems increasingly central to research. Without deliberate investment, these gaps will deepen structural inequalities in who produces knowledge and whose questions get asked • Transparency, accountability, and human oversight: As AI is adopted across the research lifecycle, from literature synthesis to hypothesis generation to data analysis, questions of reproducibility, traceability, and human accountability for scientific claims become urgent governance issues • Open-source software, open data and open AI models: Openness is a practical lever for urgent action as they can speed up transparent research, independent scrutiny, and capacity-building in AI governance. It also fits the ISC's call for active engagement by putting scientists and other stakeholders into the global dialogue so they can help shape safer, more interoperable, and publicly beneficial AI systems rather than leaving those choices to governments and industry alone.

• The impact of AI on science as a system is a critical gap in current governance discussions. AI is typically treated as a tool within science rather than as a force reshaping science as a system, affecting how knowledge is produced, validated, communicated, and used • Key concerns include: o The concentration of AI-driven research capacity in a small number of institutions and countries, deepening existing inequalities in global knowledge production o The shift from human-led to AI-assisted or AI-led research, and associated implications for scientific agency, accountability, and the integrity of the knowledge base o The risk that AI systems trained predominantly on data and epistemologies from the Global North will systematically marginalise other knowledge traditions and research priorities o The implications for science advice to governments: if AI shapes what evidence exists and how it is synthesised, it also shapes the foundation of evidence-based policymaking o The erosion of public trust in science if AI use in research is not governed transparently

• The scientific community currently lacks governance frameworks specifically designed for AI use in research as most existing guidance addresses commercial or general-purpose AI, not the full research lifecycle • Without standards for AI disclosure in research outputs, reproducibility and research integrity are at risk; the ISC is actively developing an international framework to address this gap • Researchers and institutions in the Global South face compounding disadvantages: limited access to frontier AI tools, underrepresentation in training data, and limited capacity to develop or adapt AI systems for local research priorities • Science funding agencies and national academies lack the policy guidance needed to govern AI adoption in their institutions as there is no internationally agreed baseline • Governance gaps also create risks for science advice: if AI-synthesised evidence is used in policy processes without adequate transparency standards, the reliability of that advice cannot be assured • The opportunity is significant: well-designed governance could accelerate scientific progress equitably, improve access to research infrastructure, and support the development of AI systems that serve a broader range of scientific questions and communities

• The Dialogue can serve as the primary multilateral forum for ensuring that the international science system has a recognised seat at the AI governance table as a stakeholder with specific interests and governance needs. • It can facilitate the development of shared norms and standards for AI use in publicly funded research, building on existing open science frameworks • It can provide a mechanism for aligning national and regional AI governance approaches in ways that protect the integrity of the global scientific record • It should actively counteract the fragmentation of AI governance into geopolitical blocs, which risks fragmenting the scientific commons along the same lines

• The AI Dialogue should build on existing harmonisation and governance efforts already under way in research publishing, research integrity, and broader AI-for-good ecosystems. Its added value would be to connect these strands, reduce fragmentation, and turn them into practical guidance that research actors can adopt consistently. • Initiatives to build on o The emerging Global Reporting Standard for AI Disclosure in Research, being developed with ISC, COPE, STM, and GYA, provides a concrete foundation for common disclosure practices across disciplines and publication cultures. o The World Conference on Research Integrity focus track offers a ready-made venue for convening publishers, researchers, and institutions around disclosure norms and implementation. o Broader multi-stakeholder initiatives such as AI for Good and inclusive AI coalitions show the value of linking technical, policy, and societal perspectives in a shared dialogue. o National and regional policy efforts on AI in research ecosystems can provide test beds for implementation, feedback, and adaptation across different contexts. • Added value of the AI Dialogue o It can act as a bridge between global principles and practical uptake by translating high-level standards into usable guidance for funders, universities, publishers, and researchers. o It can create a common space to align disclosure, integrity, and capacity-building agendas, avoiding duplicated work and inconsistent rules across sectors. o It can surface regional differences and implementation challenges, helping ensure the standard is globally relevant rather than shaped only by a few settings. o It can strengthen coordination across existing networks, accelerating shared learning and supporting responsible, transparent AI use in research.

• Scientific organisations should be recognised as a distinct stakeholder category with formal participation rights, not subsumed under "civil society" • Research funders (national and international) are a critical governance actor as their grant conditions and disclosure requirements can set de facto global standards • Structured working groups on science-specific AI governance issues should be established as part of the Dialogue process, with outputs feeding into plenary deliberations • The Dialogue should distinguish between consultation and co-design: affected communities (i.e., including researchers, science communicators, and scientific publishers) should be involved in shaping governance frameworks, not only invited to respond to them • Written input processes should be complemented by regional preparatory dialogues to ensure that perspectives from underrepresented regions are integrated before, not after, drafting begins

• Scientific communities from the Global South are systematically underrepresented both as AI developers and as participants in governance processes • Early-career researchers who will live with the long-term consequences of current governance decisions are rarely present in these discussions • Researchers in disciplines beyond computer science and economics such as social scientists, historians of science, philosophers of knowledge, and domain scientists, bring essential perspectives on what AI means for different fields and knowledge traditions • Indigenous knowledge communities whose epistemologies and data are at risk of appropriation or marginalisation by AI systems trained without their consent • Inclusion mechanisms should go beyond travel grants: regional knowledge hubs, translated materials, and asynchronous participation formats are needed to make engagement substantive rather than symbolic

• Structured futures dialogues: facilitated scenarios exploring long-term implications of AI for science and knowledge production, designed to surface risks and opportunities that near-term risk frameworks miss • Regional science-governance interfaces: dedicated convenings that bring together AI governance actors and the scientific community at regional level, feeding into global processes rather than receiving outputs from them • Living policy labs: ongoing, iterative engagement formats (rather than one-off consultations) where governance frameworks are tested, refined, and updated in dialogue with practitioners • Cross-sectoral scenario exercises: bringing together AI developers, scientific publishers, research funders, and policymakers to work through concrete governance dilemmas, producing actionable recommendations rather than general principles

• AI disclosure in research initiative: developing a global standard for reporting AI use in scientific outputs, addressing a concrete governance gap at the research-integrity interface • FAIR data principles: an international community-developed standard for research data governance; a model for how the science community can develop its own governance norms that then inform broader policy